Greedy compensatory agent based day-ahead photovoltaic power forecasting with a simplified deep Q-network

As the energy transition accelerates, accurate forecasting of photovoltaic (PV) power generation is crucial for the stable operation and effective management of electrical grids. However, the intermittent and variable nature of photovoltaic (PV) power output, influenced by weather conditions and environmental factors, places the PV power forecasting environment in a state of dynamic change, posing significant challenges. To address this issue, this study proposes a hybrid PV power forecasting method that combines a multi-layer perceptron (MLP) and a simplified deep Q-network (SDQN) with a greedy compensatory agent (GcaPF). The MLP provides the base forecast, while the SDQN model refines the prediction by seeking to derive the optimal compensatory value for MLP's output at each time step. The use of the ratio between actual PV power and forecasted solar irradiance as the MLP's learning target can partially mitigate the impacts of irradiance prediction errors, and the adapted SDQN architecture focuses on compensating for the prediction errors at each forecasting point through reinforcement learning (RL), leading to a more optimized and reliable forecasting outcome compared to standalone models. The average MSE reduction is 24.08 % and 23.64 % for plant1, and 10.28 % and 11.07 % for plant2 at different timesteps.